TRUST-SCORE DEVICE-AWARE INTRUSION DETECTION FOR IOT NETWORKS: DATA POISONING ATTACK ANALYSIS AND MITIGATION

This thesis analyzes the vulnerability of Machine Learning-based Intrusion Detection Systems to data poisoning attacks in IoT environments. Through label flipping experiments on the CIC IoT-DIAD 2024 dataset, the impact on two representative models, a supervised Random Forest and an unsupervised Autoencoder, is quantified. Existing defenses (data sanitization, RONI, adversarial training) are evaluated and shown to be ineffective against targeted device-level poisoning. A novel trust-score mechanism is proposed, which assesses the reliability of each source device based on the stability of its network behavior over time.